Scan to download the Future Doctor app
Medical AGI.
It starts with
one person.
From clinical reasoning to care in daily life. Under doctor governance, intelligence connects decisions, actions and outcomes around one person.
From a sound decision
to care that follows through.
Think the decision through.
Help the next step happen.
Bring the results back.
A way through
complexity.
MedGPT · The clinical strategy engine
Understand the person. Define the goal.
Begin with the patient’s condition, treatment history, practical constraints and hopes for daily life. Establish the decision, the outcomes that matter and the risks they cannot accept.
Give every option a foundation.
Develop options into strategies with conditions and branches. Weigh supporting and opposing evidence, establish where it applies, then compare benefits, risks and burdens.
Challenge. Repair. Reconsider.
Seek weak assumptions, overlooked risks and evidence against a strategy. Turn each challenge into a repair, a request for evidence, or a recorded reason to exclude and reconsider.
What could change the decision?
Identify the unknowns that could change the choice. Weigh the time, risk and burden of finding out—and specify how different results would change the next step.
Bring the decision into focus.
Bring the current preference, alternatives, evidence, disputes and reassessment conditions to the doctor. Make the reasons for acting, verifying or maintaining a plan clear.
EvidenceOS brings supporting evidence, counterevidence and applicability into each choice.
01 / 04
Clinical judgment.
Built into
every step.
Medex · The doctor’s decision workspace
Correct a fact. Question the evidence. Reframe the goal. Review a strategy.
A doctor’s revision becomes a change to the decision itself. The affected reasoning is reconsidered, with the reason for each change kept visible.
Try Future Doctor Medex“The patient would prefer fewer hospital stays.”
“The patient would prefer fewer hospital stays.”
“The patient would prefer fewer hospital stays.”
“The patient would prefer fewer hospital stays.”
“The patient would prefer fewer hospital stays.”
Future Doctor · The patient experience
A decision becomes
a journey of care.
The same clinical question connects the doctor’s workspace and the patient’s daily life. Every step has a person responsible. Every result has a way back. Explore patient services
Find a way forward.
Start with the question that matters to the patient. MedGPT organizes evidence and options. In Medex, doctors independently review, revise and sign opinions within their scope. The patient experience makes the reasoning, disagreements and next step understandable.
Bring evidence and clinical judgment to the next visit.
- Current goal
- Clarify the next step and reduce repeated visits
- Opinion & scope
- Reasons, disagreements and conditions are visible
- Clinical action
- The treating clinician has not yet accepted the handoff
The next action still requires a real consultation and clinical confirmation.
Make the next step happen.
The patient has one path manager to turn to. The system organizes records, tasks and handoffs; the manager coordinates the service. Authorized clinicians and care institutions accept and carry out clinical actions, with a time and a fallback for each step.
The next visit. An accepted handoff. A clear update.
- Service coordination
- Path manager · Accepted
- Clinical responsibility
- Treating clinician · Scope of this visit accepted
- Next action
- Attend the confirmed visit and record what took place
- Next update
- At the agreed time after this visit
- If delayed
- Manager coordinates; a clinician reviews timing implications
After scheduling and acceptance, the service verifies what actually happened.
Let reality inform the next decision.
Collect feedback and new records, verify their sources and timing, and recheck the assumptions behind the plan. An authorized clinician interprets the results against the agreed goal. New facts enter PatientRepo and reopen only the reasoning they affect.
The records are back. The work continues.
- Result interpretation
- Responsible clinician · Accepted, not yet completed
- Current state
- Source and timing verified; goal attainment to be assessed
- Possible next step
- Verify attainment, observe safely, reopen a decision or transfer with acceptance
An uploaded result still needs interpretation and assessment against the agreed endpoint.
One care journey.
Two rhythms of support.
Every decision has a history.
And a next chapter.
PatientRepo
What we knew. Why we chose. What happened next. A shared, living context for doctors and
patients throughout care.
Record what was known and why a choice was made. Give the next decision a foundation.
Compare new outcomes with the observation criteria set in advance. Separate what happened from what still needs to be tested.
As care progresses, circumstances and hopes can change. Goals confirmed by the doctor inform the next trade-off.
Once new information is supplied and verified, trace its impact through dependencies. Reopen the relevant reasoning and preserve the basis of each decision.
One person. A continuous system.
Think it through.
Carry it forward.
Learn what happened.
Our ambition: help a critical medical question reach its next lawful, verifiable, better state—with less burden on the patient.
As clinical problems change, the system must reorganize evidence, options, verification and collaboration. Reasoning, doctor governance and feedback over time make up the architecture. Decision, action and outcome are how its value reaches the patient.
Results must be interpreted against the agreed goal. “Better” can mean safe observation, no intervention, or a transfer that someone has accepted. This is the value the architecture aims to deliver.
CLINICAL RESEARCH · MedGPT
A vision, put to
a clinical test.
Future Doctor and the thoracic surgery team at Peking Union Medical College Hospital jointly studied MedGPT in offline evaluations of complex lung cancer cases—testing its ability to organize strategies and support doctors’ plans.
Doctors, all from tertiary Grade A hospitals
Complex lung cancer cases across two evaluation sets
Doctor responses across three assistance conditions
Better clinical plans.
Adjusted APAS rose from 53.40 to 66.27: a 12.87-point gain. Compared with RAG reference materials, the relative gain was 8.6%.
250 doctors · 750 responses per condition. Each doctor answered different case tasks across conditions.
Model capability, organized into strategy.
On the same 40 cases, the full MedGPT workflow scored 34.8%–76.2% higher than the three Direct generation baselines, averaged across three AI judges.
Direct means generation outside the MedGPT workflow. This comparison does not cover every agent architecture or clinical use of these models.
Experts preferred the MedGPT plan.
Against the treating clinicians’ original plans, MedGPT won 16 of 20 comparison modules: 80.0%. There were 4 losses and no ties.
These are comparison modules, not numbers of experts or patients. This is a separate evaluation branch from the 250-doctor assistance experiment.
Study design, measures and interpretation
Cases and reference rules
40 reference cases supported rule development, doctor tasks and system comparisons; a separate 60-case set supported source-masked comparisons. Twenty senior chief physicians across six disciplines established the rules before reviewing the outputs. The 240 case-level rules include 40 safety rules.
Doctor assistance experiment
Each of the 250 doctors completed nine different case tasks: three unaided, three with RAG references and three with a MedGPT strategy review package. The doctors came from 98 normalized institutional centers across 27 province-level regions; these are not 98 patient-enrollment trial centers.
What the scores measure
APAS is a 0–100 plan-quality score based on weighted compliance with expert acceptable-pathway rules. CSPR is the full-rule pass rate under the frozen joint-gating rules; the reported adjusted probabilities cannot be converted into raw pass counts. For MedGPT assistance vs unaided / RAG, APAS P values were 8.49 × 10⁻⁵¹ / 1.98 × 10⁻⁹.
Replication and expert comparisons
Three AI judges each scored the same responses three times; all nine runs favored MedGPT assistance. Two doctors reviewed a 120-response sample in the same direction. Across all 120 source-masked comparison modules from 60 cases, MedGPT recorded 72 wins, 40 losses and 8 ties: 64.3% of non-ties. The three evidence branches above are not a pooled win rate.
The results provide initial evidence for MedGPT’s approach to organizing complex clinical decisions. They evaluate written plans and rule compliance on fixed lung cancer cases. Patient outcomes, generalization across diseases and the complete MedGPT + Medex + PatientRepo architecture still require further study.
Study results · September 2026